Everyone Just Agreed on the Foundation. Nobody's Answering Who Gets to Run It.

The whole data industry just agreed on what good looks like.

This week Fivetran and dbt Labs closed their merger. The deal was announced back in October, and on June 1 it became official: one company, with George Fraser as CEO and Tristan Handy as President, serving more than 100,000 data teams. They called it "the data infrastructure for trusted AI agents."

Read past the headline and the thesis is clear. Trust in AI doesn't get built in the model. It gets built underneath the model, at the infrastructure layer: on open standards, on governed business logic, and on high-quality data an agent can actually rely on. Tristan Handy put it plainly. "Trust is built at the infrastructure layer, on high-quality tooling and on open standards. That's the bet we're making together."

I agree with almost all of it. It's the same bet I made when I started building OptimaFlo: open formats, data that stays in your own cloud, and a governed semantic layer underneath everything. When the two biggest names in the modern data stack merge and land on the same foundation you've been building on, that's not a threat. That's validation.

But there's one line in the announcement I don't agree with. And I think it's the most important line in the whole thing.

"No human in the loop"

Here's the exact passage:

"And many organizations want to move into a world where most agents are autonomous — no human in the loop."

This is where it gets messy.

A clean, governed foundation is necessary. It is not sufficient. Governance and guardrails don't decide which joins are right for your business. They don't know that "active customer" means something different in your finance reports than it does in your product analytics. They don't know which of three revenue tables is the one leadership actually trusts. That context lives in people's heads, in Slack threads, in the way your company actually operates. It is not in the schema.

And LLMs are still hard to make deterministic. Ask the same question twice, get two different queries. Most days that's fine. The day it isn't, you don't get an error. You get a confident answer that's quietly wrong.

So play out the autonomous version. An agent runs a join that looks reasonable and isn't. It doesn't have the full picture of what the business is trying to do. There's no one to catch it before the number ships. Now that number is in a board deck, or worse, in a pricing decision. The agent was fast. It was governed. It was also wrong, and nobody knew until the damage was done.

"Trusted" and "no human in the loop" are not the same thing. Treating them as the same is how you build something that feels trustworthy right up until it costs you.

I think AI agents absolutely belong in data engineering. I've built my whole platform around them. But a human in the loop isn't friction to be optimized away. It's the quality gate. Someone has to direct the work and approve what ships, especially while the models are still this unpredictable.

The question nobody's answering: who actually runs this?

Look at who the giants build for. The merger announcement names OpenAI, LVMH, Pfizer, Verizon, Siemens. More than 100,000 data teams.

The operative word is teams.

This foundation, the one everyone now agrees is correct, was designed to be operated by people who do data for a living. Data engineers to wire it up. Analytics engineers to model it. Platform owners to keep it governed. If you're a Fortune 500 with a department for this, the new world is great.

Now picture the company that doesn't have that. A growing mid-market business. Real revenue, real customers, data spread across Stripe and HubSpot and a warehouse someone set up two years ago. They have every data problem an enterprise has. What they don't have is a data team, and they can't justify stitching together six to ten specialized tools and hiring the engineers to run them.

These companies are not a niche. They are most companies. And they are the ones this consolidation quietly leaves behind. The foundation is open and excellent. The cost of operating it is still a full team.

That gap is the whole reason OptimaFlo exists.

Same foundation. One person to run it.

OptimaFlo is built on the exact open, governed foundation the industry is converging on. Apache Iceberg and open formats, so there's no vendor lock-in. Bring-your-own-cloud, so your data never leaves your environment. A governed semantic layer and business logic underneath everything, in the platform from day one.

The difference is who operates it. Instead of a team, we give one data owner an AI data team to run the whole thing. The agents do the heavy lifting: generating pipelines, writing the SQL, modeling the layers, surfacing anomalies. The owner stays in control. They direct the work. They approve the SQL before it runs. Nothing ships on the agent's say-so alone.

That's the human in the loop, built into the product on purpose. The AI amplifies the person who owns the data. It doesn't replace them, and it doesn't pretend their judgment is optional. The work that used to take a whole team, one person can now direct, with the agents handling the volume and the human owning the call.

It's the same answer to the trust problem the giants are reaching for. We just don't think trust means taking the human out. We think it means giving one human the reach of ten, and keeping their hand on the approval.

Where this leaves us

The industry has settled the hard architectural argument. Open standards. Governed business logic. High-quality data an agent can rely on. That part is done, and the Fivetran and dbt merger is the clearest signal yet.

The open questions are the ones about people. Who gets to operate this foundation, not just the enterprises with a department for it? And does "trusted" really mean "no human in the loop"?

I don't think it does. I think the companies that get AI in data right will be the ones who kept a person in the room, directing the agents and approving what ships, all the way down.

If your team didn't have to fight the data stack to get a trustworthy answer, what would you actually build with that time?